Borrowing Boundary Calibration Across Recording Sites

Sound event detectors are asked to say not just what happened in a recording but when it started and stopped, often to within tens of milliseconds. They are also increasingly deployed as networks of recorders spread across landscapes, and only a handful of those sites ever get annotated. Two facts from the literature collide here. Detectors tuned for precise timing are fragile when the background changes. On purely non-target background audio, systems scored for fine timing produce more than a thousand spurious "ghost" events (Ronchini & Serizel). In a geographically sparse deployment, however, the settings that control this fragility (decision thresholds, allowed event durations, which background sounds count as negatives) can only be tuned at the few labeled sites. Every other site inherits settings calibrated for somebody else's wind, insects, and machinery.

This note proposes to treat boundary calibration as something to transfer between sites. For an unlabeled recording site, we describe its background using only its own unlabeled audio. We then retrieve the labeled sites whose backgrounds look most similar and borrow three things from them: their decision thresholds, their event-duration and transition statistics, and their hard negative examples. The empirical question at the center is which notion of "similar site" predicts the right calibration: acoustic background similarity, geographic distance, or satellite-derived land cover. We expect the answer to be background similarity. If so, the practical lesson for sparse networks is to label for background diversity rather than for spatial coverage.

Core Claim

The site-specific part of boundary error is predictable from a label-free description of the site's non-target background. At an unlabeled site, the oracle per-site settings are the threshold and duration parameters one would pick given that site's labels. We claim these can be regressed from the background descriptor far better than from coordinates or land cover. We further claim that transferring calibration by background similarity recovers a substantial share of the gap between global settings and oracle per-site settings, both on event-based F1 and on ghost-detection rate.

This differs from the closest work in specific ways. Sat2Sound maps satellite imagery to 10-second ambient soundscapes for unmonitored locations, and Soundscape Connectomes reconstructs masked recorder sites from graph neighbors using counts of sound clusters aggregated over ten days. Both fill geographic gaps, but at the level of static summaries, with no onsets, offsets, or thresholds. The Surrey few-shot system adapts to background within a single file, through transductive inference and negatives drawn from the same file, and never shares information across sites. Mind the Domain Gap documents that detectors collapse when moved across habitats but leaves the shift unmodeled. Our proposal models that shift explicitly, at the level of the quantities that govern boundaries, and tests whether geography carries any information about it.

Research Program

The detector itself is not the contribution; we hold it fixed. For fully supervised detection we use an onset-conditioned multi-task network in the style of Pankajakshan et al.. This is a model whose onset head anchors activity and class predictions, and which raised event F-score on URBAN-SED from 7.34% to 23.20%. We also use the duration-aware BEATS-CRNN-HSMM of Yoshinaga et al.. A hidden semi-Markov model (HSMM) is a sequence model that assigns explicit probabilities to how long an event lasts, so its duration and transition parameters are natural calibration targets. For few-shot bioacoustics we use the Surrey prototypical-network pipeline. That pipeline builds a class "prototype" from five examples and classifies frames by distance to it.

The first component is the background descriptor. For each site, we collect embeddings of low-activity frames from its unlabeled audio, identified with a class-agnostic activity detector. We summarize them as a distribution, using mean, covariance, and a small codebook histogram. Distance between sites is a distribution distance between these summaries.

The second component is calibration transfer. Let $\theta_s$ denote the calibration of labeled site $s$: per-class thresholds, HSMM duration parameters, and onset-head threshold, all tuned on that site's labels. Let $b_s$ be its background descriptor. For a target site $t$ we set

$$\hat\theta_t = \sum_{s \in \mathcal{L}} w_{ts}\,\theta_s, \qquad w_{ts} \propto \exp\!\left(-d(b_t, b_s)/\tau\right),$$

where $\mathcal{L}$ is the set of labeled sites, $d$ is the background distance, and $\tau$ is a temperature chosen by leave-one-site-out validation on labeled sites only. The same weights determine which labeled sites contribute hard negatives to the target's negative pool, alongside the target's own low-activity frames. A second variant replaces the weighted average with a small regressor from $b$ to $\theta$ trained across labeled sites. This lets it extrapolate to backgrounds that no single neighbor matches.

The third component is the diagnostic that makes the geographic question falsifiable. Across all sites with labels, we compute oracle $\theta_s$. We then measure how well it is predicted from the background descriptor, from geographic distance by inverse-distance weighting (IDW, which averages neighbors' values weighted by closeness), and from land-cover features, alone and in combination. This mirrors the node-removal test of Soundscape Connectomes, but the reconstruction target is boundary calibration rather than sonotype counts.

Baselines

The naive reference is pooled calibration. We use one set of thresholds and duration parameters tuned on all labeled sites together and apply it everywhere. This is what current practice amounts to. It is the point against which recovered gap is measured.

The strongest existing method is per-file transductive adaptation with in-file negatives, as in the Surrey system. It reports an F-measure of 68.74% on the DCASE 2022 Task 5 validation set, and transductive learning alone added 11.58 F1 points absolute according to Liang et al.. It is the strongest competitor because it already exploits the target recording's own background, which is exactly the information our descriptor uses. We run it with the identical backbone, identical five-shot supports, and the same tuning budget. We then run our transfer on top of it. Beating it means a statistically reliable gain in event-based F1 at held-out sites and a lower ghost count, not merely matching it. For supervised domestic sound detection, the reference points are Confidence Regularized Entropy, which reports event-F1 51.11 on DESED (Choi & Chang), and the HSMM system, for which there is no reported numeric result on Lacuna. We retrain both on our site splits rather than quoting their numbers.

The geographic baseline interpolates oracle calibration from labeled sites by IDW, or by land-cover similarity using satellite-derived features. The upper bound is oracle per-site calibration using the target site's own labels.

The key ablation isolates which ingredient matters. Holding everything else fixed, we compare retrieval by background similarity, by geography, by land cover, and at random. We also separate threshold transfer from negative-pool transfer from duration transfer.

Why not just…?

Why not just rely on transductive adaptation within each file? It adapts the representation to the test file. However, the threshold and duration settings still come from wherever the model was tuned, and the ghost-detection problem is a threshold problem as much as a representation problem. The comparison is direct because we stack transfer on top of transduction. If the stacked system shows no gain at held-out sites, then transduction already captures what background matching offers, and we will report that.

Why not just add coordinates or a satellite embedding as an input to the detector? This is the real threat. Sat2Sound's retrieval recall jumps from 0.534 to 0.871 when location and time metadata are fused, so location clearly carries acoustic information. Our diagnostic answers it directly. If land cover or coordinates predict oracle calibration as well as the background descriptor does, the geographic route wins, and it is cheaper for sites never recorded at all. We will also train a variant conditioned on a satellite embedding. Our prediction is that it helps only where background descriptors are unavailable, but the experiment decides.

Why not just interpolate detections spatially? Kriging-style smoothing of event outputs spreads each site's boundary errors to its neighbors. It also pulls sparse regions toward the network mean, which is the false-negative failure seen in spatial epidemiology (Colah). The IDW baseline, applied to calibration rather than to detections, is the fairer version of this idea, and it is included.

Why not just use a bigger pretrained model? Adding large pretrained audio models improved tagging but not timing in DCASE Task 4 (Ronchini et al.). A larger backbone therefore changes the representation without removing the need for site-specific calibration. We will nonetheless repeat the main comparison on a BEATS backbone.

Proposed Experiments

The first experiment is a controlled testbed. It builds synthetic "sites" with Scaper by pairing DESED and URBAN-SED foreground events with distinct background pools and signal-to-noise levels, following the TNTSNR design of Ronchini & Serizel. Ground-truth boundaries are exact here, and the background shift is known. We hold out sites and report event-based F1, PSDS scenario 1 (the timing-sensitive score), median onset and offset error in milliseconds, and ghost counts on the purely non-target synth_ntg-style clips. We expect background-matched transfer to recover most of the pooled-to-oracle gap here. If it fails even in this setting, the idea is wrong.

The second experiment uses real data from DCASE Task 5. We treat each source subset, which differ in recorder, habitat, and taxon, as a site and run leave-one-site-out evaluation. The references are the Surrey figure of 68.74% on the 2022 validation set and the 52.14% DCASE 2024 baseline reported by Liang et al. The marine and deer subsets that Liang et al. identified as hardest are the decisive cases.

The third experiment is a geographic check on the Puerto Wilches network, the 17 sites used in Soundscape Connectomes. It requires new annotation. We will add strong labels for a few frequent call types at a subset of sites, then run the diagnostic regression of oracle calibration against background, distance, and land cover. Because these sites sit within about 9 km² across oil palm, forest, and grassland, they separate geographic proximity from habitat similarity. That separation is exactly what distinguishes our hypothesis from its geographic rival.

The decisive result would be the following pattern at held-out sites: background-matched transfer beats both pooled calibration and transduction alone, geographic IDW does not, and the background descriptor explains substantially more variance in oracle thresholds than distance or land cover. If geography matches background similarity instead, the conclusion reverses, and the proposal becomes a case for satellite priors in sparse acoustic networks.